Claude Code's team automates most code review via pre-PR verification, AI and security reviewers that block merges, risk-tiered human approval, merge queues, feature flags, and fast rollback. Humans review risky code and architecture, minimizing time without raising reverts or incidents.

Datadog Security Labs evaluated Jev's prompt injection detection and shows how counterfeit classification policies can steer its decisions and bypass LangChain's tool-risk gating.

Datadog tested Sonnet 5, Composer 2.5, and GPT 5.5 with CodeGuard security rules. Only GPT 5.5 closed the cross-user document IDOR. Sonnet 5 kept injection hygiene but retained authorization gaps; Composer 2.5 kept hardcoded secret fallbacks and post-logout token validity.

What models propose after a cloud compromise, and the detections we can build from it.

Trail of Bits released patching-bugs and regression-hunt, agent plugins for reproducing bugs, developing fixes, and detecting regressions via live PoCs. regression-hunt yields 25% fewer false positives than a baseline Codex prompt, but confidently reports fewer known regressions.

Implement a ChatGPT-like LLM in PyTorch from scratch, step by step.

NCSC chapter on AI in cyber security covers autonomous cyber defence, AI system security, and national-scale resilience. Agentic AI offers defenders scale but adds risks, as threat actors already use AI.

LiteLLM 1.107.0 on PyPI and Docker images was compromised via a lookalike Berri-AI/litellm build source. Python startup hooks (cook.pth, sitecustomize.py) steal cloud, developer and AI tool credentials, exfiltrating them. A GitHub-hosted file controls a remote Python loader.

Google Threat Intelligence Group and Mandiant attribute zero-day exploitation of BeyondTrust Remote Support CVE-2026-102764 to PRC-nexus ARMORY CASTLE. The actor deployed SAWFLY webshell and BARECLAW SOCKS5 tunneler, escalated to root via SUID shell, harvested vault credentials, and deleted logs.

A modular AWS privilege escalation exploitation framework with a Metasploit-style REPL, a scriptable CLI, 80+ exploit modules, and interchangeable payloads. You can also check out the companion blog post.

A community-driven dependency manager for AI agents. Think package.json, requirements.txt, or Cargo.toml, but for AI agent configuration.

Sandboxed Platform for Agentic Code Execution: pluggable-runtime sandboxes for running agent and user code in isolation.

A benchmark for vulnerability-driven dependency upgrades in real, runnable repositories.

Amazon ECS Gateway is a new account-level resource that automatically provisions ingress infrastructure for ECS services. Platform teams define networking, TLS, and WAF settings once, and application teams opt in with a gateway connector.

Google Cloud added folder membership to VPC Service Controls perimeters, so contained projects inherit policy. Ingress/egress rules now accept folders, organizations, and Private Service Connect endpoints.